Attribute-Driven Spontaneous Motion in Unpaired Image Translation

解决的问题

前人的做法:

Ÿ   Success of image translation methods mostly imposes the requirement of working on aligned or similar domains for texture or appearance transform.

Ÿ   The building blocks of these networks, such as convolution/deconvolution layers and activation functions, are spatially corresponding.

问题:

Ÿ   As shown in Fig. 1(a-d), artifacts or ghosting could appear when nonsmile and smile faces are not geometrically aligned in image space.

Attribute-Driven Spontaneous Motion in Unpaired Image Translation

大白话解释一段我的理解:

       非成对图像翻译,比如非笑脸图(小明,男)要翻译成笑脸图(小明,男)。

       输入网络的图,非笑脸图(小明,男)和笑脸图(大红唇的小红,女),经过一系列的convolution/deconvolution layers and activation functions,得到小明的五官信息和大红唇小红的笑信息(maybe 嘴角上扬图),重组以后得到,大红唇的笑脸小明。

       结果图中的大红唇可以看作artifacts or ghosting。

期望的效果:

 Attribute-Driven Spontaneous Motion in Unpaired Image Translation

它的具体做法:

 Attribute-Driven Spontaneous Motion in Unpaired Image Translation

SPM

 Attribute-Driven Spontaneous Motion in Unpaired Image Translation

RM

12个平行采样的残差块,对生成结果进行精细化。Attribute-Driven Spontaneous Motion in Unpaired Image Translation

Mask

Ÿ   3 deconvolutional layers to up-sample f into a 1-channel mask m, the same size as the input.

Ÿ   Sigmoid layer is used as the final activation layer to range the output mask in [0; 1].

Ÿ   a regularization term Lm to enforce sparsity of masks in L1-norm: 

实验效果展示

实验详情

Ÿ   数据库:

  • CelebA  (200K 名人图像,40种属性,图像大小218*178)
  • CelebA-HQ (图像大小1024*1024)
  • RaFD (67个人,8种表情,3种眼神方向,5种拍摄角度)

选择的属性“Smiling’, ‘Arched eyebrow’, ‘Big Nose’, ‘Pointy nose’”,只训练了正脸。

Ÿ   Pytorch

Ÿ   TITAN Xp

Ÿ   训练策略(two-stage training)

ü  Adam,learning rate 1e-4

ü   LR framework, 128 *128 images, batch size 16,  iterations.

ü   Higher resolutions 256*256 or 512*512, batch size 8 , another iterations.

 Attribute-Driven Spontaneous Motion in Unpaired Image Translation

 Attribute-Driven Spontaneous Motion in Unpaired Image Translation

上一篇:使用MINST数据集训练模型并动态绘制Loss和准确率的趋势图


下一篇:【37】池化层讲解(Pooling layers)